Week 18 — Inference serving (light)¶

At a glance

Phase 5: Scale & Depth · Starts: Mon 25 Jan 2027 · Planned: 11h 00m · Light week

Build: Capstone: self-hosted vLLM (or Ollama) serving experiment with quantised model behind LiteLLM

By day¶

System Design · 30 min · Leases, fencing tokens, clock pitfalls — Kleppmann lectures on distributed locks → Consensus: Raft, leases, fencing · resource w18-sd-1
DSA · 15 min · Edit Distance → Dynamic programming 2-D · resource w18-dsa-1
Python · 15 min · Rep: GIL internals + what changes without it → CPython internals: bytecode, GIL, memory · resource w18-py-1
Communication · 20 min · Grammar: Review: hardest patterns from your error log → Week 18 drills · day 1 w18-comm-1

Agentic AI · 40 min · Inference serving: continuous batching, PagedAttention, prefix caching — vLLM vs SGLang vs Ollama → Inference serving: vLLM, SGLang, Ollama, quantization · resource w18-ai-1
DSA · 20 min · Longest Increasing Path in a Matrix → Dynamic programming 2-D · resource w18-dsa-2
Communication · 20 min · Vocabulary: Review, spaced repetition round → Week 18 drills · day 2 w18-comm-2

DSA · 15 min · Distinct Subsequences → Dynamic programming 2-D · resource w18-dsa-3
Architecture · 30 min · Data architecture: CDC, lakehouse, data mesh — keeping a vector index in sync with source systems → Data architecture: lakehouse, data mesh, CDC w18-arch-1
Python · 15 min · Rep: concurrent.futures vs asyncio.to_thread for blocking SDKs → Concurrency models: threads, processes, free-threaded 3.14t w18-py-2
Communication · 20 min · Speaking drill: Shadowing practice 3 → Week 18 drills · day 3 w18-comm-3

Agentic AI · 40 min · Quantisation (GGUF, AWQ, GPTQ, FP8) — quality/latency trade-off measured with the eval set → Inference serving: vLLM, SGLang, Ollama, quantization · resource w18-ai-2
Java/Spring AI · 20 min · Scoped values (final in 25) vs ThreadLocal for request context in virtual threads → Virtual threads, structured concurrency & scoped values · resource w18-java-1
Communication · 20 min · Idioms & phrasal verbs: Idiom review round → Week 18 drills · day 4 w18-comm-4

System Design · 30 min · Multi-region: active-passive vs active-active, RPO/RTO, failover testing → Multi-region, DR & failover w18-sd-2
DSA · 15 min · Burst Balloons → Dynamic programming 2-D · resource w18-dsa-4
Python · 10 min · Rep: profile async code with py-spy --idle → Performance & profiling w18-py-3
Communication · 20 min · Writing: Editing a colleague's draft → Week 18 drills · day 5 w18-comm-5

Agentic AI · 75 min · Capstone build: serve a quantised 7-8B model via vLLM/Ollama behind LiteLLM; route the 'classify' node to it and compare quality/cost → Inference serving: vLLM, SGLang, Ollama, quantization · resource w18-ai-3
DSA · 20 min · Regular Expression Matching → Dynamic programming 2-D · resource w18-dsa-5
Java/Spring AI · 20 min · JVM performance intro: GC choices (G1, ZGC), JFR recording of the MCP server → JVM performance, GC & Leyden/AOT w18-java-2
Communication · 20 min · Record & shadow: Shadowing practice 3 → Week 18 drills · day 6 w18-comm-6

System Design · 30 min · Written design: distributed key-value store — partitioning, replication, quorum, hinted handoff, anti-entropy (Dynamo) → Distributed key-value store · resource w18-sd-3
Architecture · 20 min · Write ADR 018 for capstone: CDC (Debezium-style) from ticket DB into the RAG index → Data architecture: lakehouse, data mesh, CDC · resource w18-arch-2
Staff+ · 20 min · Artifact: mentoring plan — 1 mentee, goals, cadence, sponsorship opportunities you'll create → Mentoring, sponsorship & growing engineers w18-staff-1
Communication · 20 min · Soft skills + weekly review: Giving upward feedback → Week 18 drills · day 7 w18-comm-7
Review · 40 min · Light-week retro + flashcards: inference serving vocabulary, quorum math → Inference serving: vLLM, SGLang, Ollama, quantization w18-rev-1